Self-Supervised Animal Detection, Tracking & Re-Identification

This paper presents a novel Multi-Object Tracking (MOT) framework primarily focusing on re-identification, diverging from traditional tracking-by-detection paradigms. We propose a deep learning-based method to process video sequences and directly infer trajectories, simplifying the tracking process and enhancing accuracy and efficiency. Our framework employs MobileNet, a lightweight convolutional neural network, to extract high-dimensional feature vectors from input images for accurate and efficient object re-identification. These features are integrated with the tracking algorithm, ensuring consistent object identification across frames. To further enhance re-identification, the system considers parameters such as velocity and distance, matching newly identified objects within a specific radius of defined pixels in the frame. This localized matching approach reduces computation and false positives by focusing comparisons near the object's expected position, adjusting the radius in real time according to the object's motion and the surrounding scene. Additionally, self-supervised learning with labeled and unlabeled data improves model generalization across various environments. At the same time, a novel occlusion-aware tracking algorithm maintains high tracking accuracy using contextual information and predictive modeling. Our approach was validated using the publicly available Dancetrack dataset and a custom dataset from Norsvin, Norway. Our system ensures continuous and accurate tracking by anticipating occlusions and maintaining continuity through virtual trajectory calculations.